Introductions and Context
Participants ranged from long-time AI experimenters (dating back to GPT-3) to more recent adopters who have shifted from simple prompt-and-copy workflows to structured, repeatable agent-based systems. Several noted that the past year, and especially the last few months, have felt materially different in terms of reliability and practical usefulness.
Case Study: Operationalizing Agents in a Consultancy
The main discussion centered on how one consultancy is using Claude CoWork, Skills, and MCP integrations to automate parts of its internal project management and sales processes.
Examples shared:
Post-meeting workflow
- Transcribe meetings
- Generate structured meeting minutes
- Extract decisions and rationale
- Update a centralized decision log
- Create and assign ClickUp tasks
Pre-meeting workflow
- Review prior meeting notes
- Check project management ticket status
- Generate draft agendas with updates
Proposal generation
- Search Slack, email, and client folders
- Identify missing information
- Generate draft proposals based on historical templates
A key architectural choice was building a foundational “client data” skill that defines where information lives and how it is structured. Other skills reference that layer, reducing duplication and improving consistency. The system is refined iteratively through meta-prompting and template adjustments.
Governance and risk were also discussed, including the lack of built-in observability in personal agent platforms and the use of third-party backups to mitigate destructive changes.
Broader Themes
Several recurring themes emerged:
Personal vs. enterprise agents Personal laptop-based agents are easier to adopt. Enterprise agent frameworks offer stronger observability and evaluation, but are harder to onboard.
Modularity over monoliths Smaller, well-scoped agent tasks (sometimes using sub-agents) tend to be more reliable than large, multi-purpose instructions.
Distribution and packaging There was discussion about how agent workflows might be delivered in the future, as skill bundles, internal marketplaces, CLI tools, or lightweight frameworks rather than traditional web apps.
Human roles Automation is increasingly handling documentation, coordination, and task generation. Participants emphasized keeping human focus on relationships, accountability, and nuanced decisions.
Tools and Updates Mentioned
Briefly discussed:
- Claude CoWork and Skills
- MCP integrations (Google Drive, Slack, calendar, transcription tools)
- Sub-agent spawning to manage context
- Remote control sessions for coding from mobile
- Scheduled tasks within agent environments
Next Steps
We plan to continue meeting on a biweekly cadence with future sessions including:
- Additional workflow demos
- Client case studies
- Governance and liability considerations
- Packaging and distributing agent systems
- Practical marketing and research workflows using AI
If you’re actively building or testing AI-driven workflows and want to compare notes with others doing similar work, you’re welcome to join the next session.